CVE-2024-5452: PyTorch Lightning RCE

CISO Take

Any self-hosted pytorch-lightning application in default configuration is vulnerable to unauthenticated remote code execution — no credentials required, no user interaction needed. The delta endpoint is enabled by default, making internet-exposed ML training servers trivially exploitable by anyone with network access. Patch to lightning >= 2.3.3 immediately; if patching is blocked, restrict or disable the delta endpoint and isolate the host from untrusted networks.

What is the risk?

Extremely high risk. CVSS 9.8 with no authentication, no user interaction, and low attack complexity means any network-reachable pytorch-lightning instance is a critical exposure. EPSS of 0.567 signals ~57% probability of exploitation in the wild within 30 days — treat this as actively exploited. ML training infrastructure is often trusted-network territory with inadequate hardening, and post-exploitation access to GPU servers, model weights, training data, and cloud credentials makes these high-value targets for IP theft, lateral movement, and cryptomining.

What systems are affected?

Package	Ecosystem	Vulnerable Range	Patched
PyTorch Lightning	pip	< 2.3.3	`2.3.3`
31.2K OpenSSF 5.9 1.6K dependents Pushed 13d ago 57% patched ~373d to patch Full package profile →
PyTorch	pip	—	No patch
100.9K OpenSSF 6.4 22.7K dependents Pushed 3d ago 11% patched ~216d to patch Full package profile →

How severe is it?

CVSS 3.1

9.8 / 10

EPSS

26.5%

chance of exploitation in 30 days

Higher than 98% of all CVEs

Source: EPSS v3 — FIRST.org

Exploitation Status

Exploit Available

Exploitation: MEDIUM

Sophistication

Moderate

Exploitation Confidence

medium

○ CISA SSVC: Public PoC

○ Public PoC indexed (trickest/cve)

○ EPSS exploit prediction: 26%

Composite signal derived from CISA KEV, VulnCheck KEV, CISA SSVC, EPSS, Metasploit, Exploit-DB, trickest/cve, Nuclei templates, and inthewild.io exploitation reports.

What is the attack surface?

AV Network

AC Low

PR None

UI None

S Unchanged

C High

I High

A High

What should I do?

6 steps

PATCH

Upgrade to lightning >= 2.3.3 immediately (pip install 'lightning>=2.3.3').
INTERIM WORKAROUND

If patching is blocked, disable the delta endpoint in application configuration or block it at the reverse proxy/WAF layer before the patch is applied.
NETWORK ISOLATION

Ensure pytorch-lightning applications are never exposed to untrusted networks; enforce strict firewall rules limiting access to authorized IP ranges and require VPN for all ML platform access.
INVENTORY

Audit all lightning/pytorch-lightning installations across your environment using package scanners (pip list, Safety, Trivy, Grype, Syft) — pay special attention to CI/CD systems and shared ML platforms.
DETECTION

Monitor for anomalous child process spawning from Python/lightning processes; alert on unexpected outbound connections from ML training hosts; review delta endpoint access logs for suspicious payloads.
CREDENTIAL ROTATION

If any instance may have been exposed, treat the host as compromised — rotate all credentials, API keys, and cloud IAM tokens present on the system.

What does CISA's SSVC say?

Decision Attend

Exploitation poc

Automatable Yes

Technical Impact total

Source: CISA Vulnrichment (SSVC v2.0). Decision based on the CISA Coordinator decision tree.

How is it classified?

Code Execution Supply Chain Framework AML.T0010.001 - AI Software AML.T0025 - Exfiltration via Cyber Means AML.T0049 - Exploit Public-Facing Application AML.T0050 - Command and Scripting Interpreter AML.T0072 - Reverse Shell

Which compliance frameworks are affected?

This CVE is relevant to:

EU AI Act

Article 15 - Accuracy, robustness and cybersecurity

ISO 42001

A.6.2 - AI system security and resilience

NIST AI RMF

MANAGE 2.2 - Mechanisms are in place to identify and address risks from third-party AI software and services

OWASP LLM Top 10

LLM05 - Supply Chain Vulnerabilities

Frequently Asked Questions

What is CVE-2024-5452?

Any self-hosted pytorch-lightning application in default configuration is vulnerable to unauthenticated remote code execution — no credentials required, no user interaction needed. The delta endpoint is enabled by default, making internet-exposed ML training servers trivially exploitable by anyone with network access. Patch to lightning >= 2.3.3 immediately; if patching is blocked, restrict or disable the delta endpoint and isolate the host from untrusted networks.

Is CVE-2024-5452 actively exploited?

Proof-of-concept exploit code is publicly available for CVE-2024-5452, increasing the risk of exploitation.

How to fix CVE-2024-5452?

1. PATCH: Upgrade to lightning >= 2.3.3 immediately (pip install 'lightning>=2.3.3'). 2. INTERIM WORKAROUND: If patching is blocked, disable the delta endpoint in application configuration or block it at the reverse proxy/WAF layer before the patch is applied. 3. NETWORK ISOLATION: Ensure pytorch-lightning applications are never exposed to untrusted networks; enforce strict firewall rules limiting access to authorized IP ranges and require VPN for all ML platform access. 4. INVENTORY: Audit all lightning/pytorch-lightning installations across your environment using package scanners (pip list, Safety, Trivy, Grype, Syft) — pay special attention to CI/CD systems and shared ML platforms. 5. DETECTION: Monitor for anomalous child process spawning from Python/lightning processes; alert on unexpected outbound connections from ML training hosts; review delta endpoint access logs for suspicious payloads. 6. CREDENTIAL ROTATION: If any instance may have been exposed, treat the host as compromised — rotate all credentials, API keys, and cloud IAM tokens present on the system.

What systems are affected by CVE-2024-5452?

This vulnerability affects the following AI/ML architecture patterns: ML training pipelines, model development environments, data science platforms, model serving.

What is the CVSS score for CVE-2024-5452?

CVE-2024-5452 has a CVSS v3.1 base score of 9.8 (CRITICAL). The EPSS exploitation probability is 26.49%.

What is the AI security impact?

Affected AI Architectures

ML training pipelinesmodel development environmentsdata science platformsmodel serving

MITRE ATLAS Techniques

AML.T0010.001 AI Software

AML.T0025 Exfiltration via Cyber Means

AML.T0049 Exploit Public-Facing Application

AML.T0050 Command and Scripting Interpreter

AML.T0072 Reverse Shell

Compliance Controls Affected

EU AI Act: Article 15

ISO 42001: A.6.2

NIST AI RMF: MANAGE 2.2

OWASP LLM Top 10: LLM05

What are the technical details?

Original Advisory

A remote code execution (RCE) vulnerability exists in the lightning-ai/pytorch-lightning library version 2.2.1 due to improper handling of deserialized user input and mismanagement of dunder attributes by the `deepdiff` library. The library uses `deepdiff.Delta` objects to modify application state based on frontend actions. However, it is possible to bypass the intended restrictions on modifying dunder attributes, allowing an attacker to construct a serialized delta that passes the deserializer whitelist and contains dunder attributes. When processed, this can be exploited to access other modules, classes, and instances, leading to arbitrary attribute write and total RCE on any self-hosted pytorch-lightning application in its default configuration, as the delta endpoint is enabled by default.

Exploitation Scenario

An adversary scans for internet-exposed pytorch-lightning applications via Shodan or Censys targeting known default ports. Upon identifying a target running lightning < 2.3.3, they craft a malicious serialized deepdiff.Delta object that embeds dunder attribute modifications (e.g., __class__, __reduce__) bypassing the deserializer whitelist. The payload is submitted as a standard POST to the delta endpoint — no authentication required. When the server processes the delta, the dunder attributes pivot through Python's object model to import os or subprocess, executing an attacker-controlled command with the privileges of the lightning process. The attacker then exfiltrates model weights, training datasets, and cloud credentials from environment variables, or drops a persistent implant for ongoing access to the ML development environment. The entire chain requires no prior knowledge of the target beyond its network address.

Weaknesses (CWE)

CWE-913 Improper Control of Dynamically-Managed Code Resources Primary CWE-913 Improper Control of Dynamically-Managed Code Resources Primary CWE-915 Improperly Controlled Modification of Dynamically-Determined Object Attributes Primary CWE-915 Improperly Controlled Modification of Dynamically-Determined Object Attributes

CWE-913 — Improper Control of Dynamically-Managed Code Resources: The product does not properly restrict reading from or writing to dynamically-managed code resources such as variables, objects, classes, attributes, functions, or executable instructions or statements.

[Implementation] For any externally-influenced input, check the input against an allowlist of acceptable values.
[Implementation, Architecture and Design] Refactor the code so that it does not need to be dynamically managed.

Source: MITRE CWE corpus.